ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33

Contents lists available at SciVerse ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing

journal homepage: www.elsevier .com/ locate/ isprs jprs

Mapping tropical forests and rubber plantations in complex landscapesby integrating PALSAR and MODIS imagery

Jinwei Dong a,⇑, Xiangming Xiao a, Sage Sheldon a, Chandrashekhar Biradar a, Guishui Xie b

a Department of Botany and Microbiology, and Center for Spatial Analysis, University of Oklahoma, 101 David L. Boren Blvd., Norman, OK 73019, USAb Rubber Research Institute, Chinese Academy of Tropical Agricultural Sciences, Baodaoxincun, Danzhou, Hainan 571737, China

a r t i c l e i n f o

Article history:Received 31 October 2011Received in revised form 15 July 2012Accepted 23 July 2012Available online 21 September 2012

Keywords:PALSARMODISEvergreen forestDeciduous forestRubber plantationHainan

0924-2716/$ - see front matter � 2012 Internationalhttp://dx.doi.org/10.1016/j.isprsjprs.2012.07.004

⇑ Corresponding author. Tel.: +1 405 325 6091.E-mail addresses: jinwei.dong@ou.edu (J. Don

(X. Xiao), sage.sheldon@ou.edu (S. Sheldon), chandra.

a b s t r a c t

Knowledge of the spatial distribution of forest types in tropical regions is important for implementationof Reducing Emissions from Deforestation and Forest Degradation (REDD), better understanding of theglobal carbon cycle, and optimal forest management. Frequent cloud cover in moist tropical regions poseschallenges for using optical images to map and monitor forests. Recently, Japan Aerospace ExplorationAgency (JAXA) released a 50 m orthorectified mosaic product from the Phased Array Type L-band Syn-thetic Aperture Radar (PALSAR) onboard the Advanced Land Observing Satellite (ALOS). PALSAR data pro-vides information about the land surface without cloud interference. In this study we use the fine beamdual (FBD) polarization PALSAR 50 m mosaic imagery and a Neural Network (NN) method to produce aland cover map in Hainan Island, China. Subsequently, forest areas are classified into evergreen anddeciduous forests and rubber plantations are mapped using vegetation and land surface water indicesderived from 250 to 500 m resolution MODIS products. The PALSAR 50 m forest cover map, MODIS-basedforest types and rubber plantation maps are fused to generate fractional maps of evergreen forest, decid-uous forest and rubber plantation within 500 m or 250 m pixels. PALSAR data perform well for land coverclassification (overall accuracy = 89% and Kappa Coefficient = 0.79) and forest identification (both theProducer’s Accuracy and User’s Accuracy are higher than 92%). The resulting land cover maps of forest,cropland, water and urban lands are consistent with the National Land Cover Dataset of China in 2005(NLCD-2005). Validation from ground truth samples indicates that the resultant rubber plantation mapis highly accurate (the overall accuracy = 85%). Overall, this study provides insight on the potential ofintegrating cloud-free 50 m PALSAR and temporal MODIS data on mapping forest types and rubber plan-tations in moist tropical regions.� 2012 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS) Published by Elsevier

B.V. All rights reserved.

1. Introduction

Tropical forests play an important role in the terrestrial carboncycle and reduce the amount of greenhouse gases such as carbondioxide (CO2), carbon monoxide (CO) and nitrogen monoxide(NO) in the atmosphere (Lelieveld et al., 2008). Tropical forests alsoprovide many ecosystem services that substantially affect humanwell-being (Foley et al., 2005; Pielke, 2005). Both human-induceddeforestation (primarily to convert land to agricultural uses) andnatural disturbance (e.g. fire, drought, wind blow-down) occurextensively in tropical regions (Bond-Lamberty et al., 2007; Kum-mer and Turner, 1994; Page et al., 2002; Sakaguchi et al., 2011).Plantations used for production of biofuels (e.g. oil palm) andindustrial resources (e.g. rubber, Hevea brasiliensis) have expanded

Society for Photogrammetry and R

g), xiangming.xiao@ou.edubiradar@ou.edu (C. Biradar).

rapidly in tropical regions in the last 50 years (Fox and Vogler,2005). This expansion has brought along a detrimental cascade ofenvironmental effects including increasing threats to biodiversityand reduction in forest carbon stocks (Li et al., 2007; Ziegleret al., 2009). Accurate information on the area and spatial distribu-tion of natural and planted forests in tropical areas is necessary forthe implementation of Reducing Emissions from Deforestation andForest Degradation (REDD) (Achard et al., 2007) and for modelingglobal carbon cycles (Dixon et al., 1994).

During the past few decades, optical remote sensing has beenwidely utilized for forest mapping (Asner et al., 2005; Collinset al., 2004; Thessler et al., 2008; Xiao et al., 2009, 2002). Previousstudies have explored the potential for tropical forest mappingusing imagery from the Advanced Very High Resolution Radiome-ter (AVHRR) (Achard and Estreguil, 1995; Achard et al., 2001),SPOT4-VEGETATION (Stibig et al., 2004; Stibig and Malingreau,2003) and Moderate Resolution Imaging Spectroradiometer(MODIS) (Miettinen et al., 2012). Most of these studies employed

emote Sensing, Inc. (ISPRS) Published by Elsevier B.V. All rights reserved.

J. Dong et al. / ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33 21

unsupervised classification due to the difficulty of ground truthsampling in tropical forest regions. Landsat Thematic Mapper(TM) images with 30 m spatial resolution and a 16-day revisit cycleare an important data source (Huang et al., 2009; Townshend andJustice, 1988); however, it is often difficult to obtain cloud-freeLandsat images in tropical regions due to frequent cloud coverand moist climate (Asner, 2001). The images from the MODIS sen-sors have been used to map forest and detect deforestation at re-gional and global scales (Friedl et al., 2002; Giri et al., 2005;Morton et al., 2005; Tottrup et al., 2007; Xiao et al., 2009). Dailyimage acquisition by the MODIS sensors partly reduces cloud prob-lems as compared to Landsat, providing valuable information toidentify and map different forest types (Friedl et al., 2002; Xiaoet al., 2009, 2002). However, its relatively coarse spatial resolution(250–1000 m) makes it difficult to accurately quantify and mapforest areas at the regional scale due to mixture of land cover typeswithin pixels. High spatial resolution remotely sensed imagery(e.g. SPOT-5, IKONOS, and aerial photographs), on the order of 1–50 m, are a very effective data source for local land use and landcover classification (Kabir et al., 2010; Perea et al., 2010; Suet al., 2010), but are not widely used in regional level monitoringdue to the high cost of image acquisition and intensive computa-tion resource requirements.

Images from synthetic aperture radar (SAR) offer an alternativedata source for mapping tropical forests (Ardila et al., 2010; Simardet al., 2000). The radar illuminates vegetation types with micro-wave energy, recording return energy that is related to above-ground biomass and structure. Longer radar wavelength L-bandSAR is better suited to the delineation of forest than other wave-lengths because of its greater penetration through the canopy

Fig. 1. The workflow

(Baghdadi et al., 2009). The Phased Array Type L-band SyntheticAperture Radar (PALSAR) data is not subject to cloud interference,making it a more effective data source for forest mapping at the re-gional scale in moist tropical regions. PALSAR is onboard the Ad-vanced Land Observing Satellite (ALOS) launched by the JapanAerospace Exploration Agency (JAXA) in January of 2006, and itprovides polarimetric radar images for most of the global land sur-face. PALSAR images have been used for many applications, includ-ing forest, crop, and ice mapping (Torbick et al., 2011; Xiao et al.,2010; Xie et al., 2010; Yang et al., 2010). The PALSAR team hasdeveloped two data products for the public: (1) the PALSAR 50 mOrthorectified Mosaic Product, and (2) the PALSAR 500 m BrowseMosaic Product. The publically released PALSAR 50 m mosaic prod-uct covers a large portion of Asia, and has recently been evaluatedfor regional forest monitoring potential in insular Southeast Asia(Longepe et al., 2011; Miettinen and Liew, 2011) with positive re-sults. However, further evaluation of the potential of PALSAR 50 mmosaic product for mapping tropical forests in many regions isneeded along with the development of new methodologies to chal-lenges in those regions with complex landscapes and land use.

Hainan Island, the most representative tropical region in China,underwent dramatic changes in land use and land cover during thepast few decades (Liu et al., 2010; Xu et al., 2002; Zhang et al.,2010). With the increasing demand for rubber products, rubberplantations continue to expand in Southern China (Qiu, 2009; Zie-gler et al., 2009). It is necessary to develop an accurate and updatedrubber distribution map for improving our understanding of landuse change and carbon and water cycles (Li and Fox, 2011, 2012).Hainan Island is now thought to have the largest area of rubberplantation in China (Chen et al., 2010). There is an increasing need

of this study.

22 J. Dong et al. / ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33

to obtain accurate information on the spatial distribution and arealextent of rubber plantations in Hainan Island.

The objective of this study was to quantify the extent and spa-tial distribution of evergreen and deciduous forests as well as rub-ber plantations on Hainan Island. PALSAR 50 m mosaic productsand MODIS time series data (MOD09A1 and MOD13Q1) acquiredin 2007 were used. This study was conducted through three com-ponents (Fig. 1): (a) creating a 50 m resolution land cover map withPALSAR mosaic data and aggregating the resultant map to produceforest percentage maps in 250 m and 500 m resolutions; (b) sepa-rating evergreen and deciduous forests as well as rubber planta-tions using a phenology-based approach with a time series ofMODIS imagery; and (c) deriving area proportion distributionmaps of evergreen, deciduous forests and rubber plantation bycombining results from the previous two stages.

2. Materials and methods

2.1. A brief description of the study area

Hainan Island is located in southern China with a geographicalarea of 34,000 km2, mostly located in a tropical area. The topogra-phy of the island is complex, characterized by hilly regions in themiddle surrounded by lowlands in the coastal regions (Fig. 2).The Wuzhi Mountain is the highest mountain with an elevationof 1867 m above sea level. Climate on the island is characterizedas a tropical monsoon climate. Annual mean temperature isapproximately 23–25 �C and monthly temperature varies between�16 �C in January to �30 �C in May to July. Annual precipitation isapproximately 1500 mm, and most of precipitation occurs be-tween May and October.

There are a variety of vegetation types on the island. Historically,most of the island was covered by natural forests (Lin and Zhang,2001); however, human exploitation has led to significant defores-tation, and the natural forest area decreased from 1.2 � 104 km2 inthe 1950s to 0.42 � 104 km2 in 1979 (Lin and Zhang, 2001). After1994, extensive reforestation occurred on the island, and forest cov-er increased to 1.84 � 104 km2 in 2004 (Hunan, 1999). Much of thenewly planted areas were cash plantations which are included in theforest classes in the land cover classification developed in this study.The cropland area is approximately 0.73 � 104 km2 and the agricul-tural intensity is high. At present, cash trees and crops, such as rub-ber, coconut, oil palm, and betel nut, are widely planted on the

Fig. 2. Spatial distribution of terrain in Hainan Island, China. The DEM data isderived from the Shuttle Radar Topography Mission (SRTM) 90 m database.

island. Human activities played an important role in these processesof deforestation, reforestation and land transformation. For exam-ple, several politico-economic activities such as the ‘‘rubber planta-tion campaign’’ in the 1950s, the ‘‘land reclamation campaign’’ in the1960s, the ‘‘crop breeding campaign’’ in the 1970s, and the ‘‘open-door economic reform’’ in the 1980s (Cai, 1994; Christian and Nar-ong, 2005) affected land use and land cover. Rubber plantations in-creased rapidly due to increasing rubber demand in recent decades,and a significant portion of natural tropical rainforest was convertedinto rubber plantations.

2.2. Data

2.2.1. PALSAR data and pre-processingThe PALSAR 50 m Orthorectified Mosaic product is provided by

JAXA and available at the ALOS Kyoto and Carbon Initiative officialwebsite (http://www.eorc.jaxa.jp/ALOS/en/kc_mosaic/kc_mo-saic.htm). It is created from images of the ascending path. The ori-ginal PALSAR images with an off-nadir angle of 34.3� wereresampled into the 50 m by 50 m mosaic. The product is producedonce a year and the dates of image acquisition vary between years.It has been geometrically rectified using 90 m digital elevationmodel (DEM) and geo-referenced into geographical latitude andlongitude coordinates (Longepe et al., 2011).

In this study, we used HH and HV polarization data from thePALSAR 50 m mosaic product acquired with Fine Beam Dual(FBD) polarization observational mode in October 2007. We con-verted the Digital Number (DN) values (amplitude values) into nor-malized radar cross section in decibel (dB), based on theconversion formula from JAXA (Rosenqvist et al., 2007) and theparameters from the metadata file:

r0ðdBÞ ¼ 10� log10DN2 þ CF ð1Þ

where r0 is the backscattering coefficient, DN is the digital numbervalue of pixels in HH or HV, and CF is the absolute calibration factor.

We calculated two additional indicators using HH and HV data(dB data): (1) the ratio of HH and HV (Ratio = HH/HV) and (2) thedifference between HH and HV (Difference = HH–HV). Both the Ra-tio and Difference images have been used in forest and forestchange monitoring (Miettinen and Liew, 2011; Wu et al., 2011).

2.2.2. MODIS data and pre-processingWe used two MODIS products: (1) the 500 m MODIS land sur-

face reflectance product (MOD09A1), which provides data in spec-tral ranges that most benefit the study of land surfaces andvegetation in an 8-day composite with quality screening; and (2)the 250 m vegetation indices product (MOD13Q1), which is de-signed to provide consistent spatial and temporal comparisons ofvegetation conditions in 16-day intervals. Detailed informationabout these products is available on the site (http://lpdaac.usgs.-gov/products/) maintained by the NASA Land Processes DistributedActive Archive Center (LP DAAC).

Three vegetation indexes, the Normalized Differential Vegeta-tion Index (NDVI) (Tucker, 1979), the Enhanced Vegetation Index(EVI) (Huete et al., 2002, 1997), and the Land Surface Water Index(LSWI) (Xiao et al., 2004, 2005), were calculated from MOD09A1data with the following formulas:

NDVI ¼ qNIR1 � qred

qNIR1 þ qredð2Þ

EVI ¼ 2:5� qNIR1 � qred

qNIR1 þ 6� qred � 7:5� qblue þ 1ð3Þ

LSWI ¼ qNIR1 � qSWIR1

qNIR1 þ qSWIR1ð4Þ

J. Dong et al. / ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33 23

where qblue, qred, qNIR1, and qSWIR1 are the reflectance values from theblue (459–479 nm), red (620–670 nm), NIR1 (841–875 nm) andSWIR1 (1628–1652 nm) bands, respectively. In this study we used500 m NDVI, LSWI and EVI data in 8-day internals and 250 m NDVIdata in 16-day internals.

2.3. Land cover map based on analysis of PALSAR 50 m mosaic data

Google Earth provides geometrically rectified images with ahigh horizontal accuracy and resolution (Potere, 2008), and hasbeen utilized in several recent land cover classification studies (Co-hen et al., 2010; Huang et al., 2010; Potere, 2008). In this study, weused Google Earth images over Hainan Island to collect a trainingdataset for land cover classification, and most of the images in Goo-gle Earth were acquired during 2006–2010, consistent with thetime of PALSAR data (2007). From these images we selected 52 re-gions of interest (ROI) for forest, 54 for cropland, 12 for water, and14 for urban land. The sizes of urban land and water ROIs were rel-atively large, as both water and urban land are distributed in largesizes and easy to extract. The ROI pixel numbers for four land cov-ers were abundant.

The feed-forward Neural Network (NN) algorithm with one hid-den layer (Richards and Jia, 2006) and PALSAR 50 m data (HH, HV,Ratio and Difference images) were used to map four land-covertypes (forest, water, urban and cropland) of Hainan Island. A num-ber of parameter compositions were tried and evaluated, and final-ly we verified and reported the parameters as following: theactivation method was set as Logistic, with parameters viz.: trainingthreshold contribution at 0.9, which determines the internal weightsand affects the activation level of the nodes; training rate at 0.2,which decides the magnitude of the adjustment of the weights;training momentum at 0.9; training RMS exit criteria at 0.1; andtraining iterations number at 2500. Training data sets were essentialfor the NN classifier, as every sample was taken into considerationin the training (Kavzoglu, 2009). We applied the ROIs mentionedabove to map land cover on the island.

Finally, the resultant 50 m PALSAR-based forest map providesmore precise information for forest canopy cover than that fromMODIS data (Fig. 6a). The PALSAR 50 m spatial resolution mapwas aggregated to calculate the percent forest canopy cover (%)within each MODIS pixel from MOD13Q1 (250 m) and MOD09A1(500 m) products. These fractional cover maps of forest were laterused together with the per-pixel binary maps of forests fromMODIS to estimate the spatial distribution and areas of evergreenand deciduous forests.

2.4. Separating evergreen and deciduous forest with MOD09A1 data

Time series LSWI data from MODIS imagery was used for map-ping evergreen forest based on its temporal profile characteristics(Xiao et al., 2009). The resultant outputs were tested against otherexisting global forest maps, such as the standard MODIS Land Cov-er data product (MOD12Q1) (Friedl et al., 2002), FAO FRA 2000, andGLC2000 (Bartholome and Belward, 2005). A MODIS pixel is as-signed to evergreen forest if it has LSWI values larger than 0 inall good-quality observations throughout a year (Xiao et al.,2009); deciduous forests do not have all LSWI values above 0 forthe entire year.

The resultant evergreen forest map from MOD09A1 imagery(Fig. 6c) was overlaid with the fractional map of forests from PAL-SAR imagery, and the total area of evergreen forest was calculatedby summing the fractions of PALSAR-based forests within thoseMODIS-based evergreen forest pixels. We then considered theremaining forest area of the PALSAR-based forest map as deciduousforests.

2.5. Mapping rubber plantations using phenology from MOD13Q1 data

When rubber was introduced from British Malaya in the 20thcentury, Hainan Island was not a suitable environment for its cul-tivation. Subsequent improvement in rubber germplasm allowed itto adapt to the frequent typhoons and low winter temperatures ofHainan Island. Rubber plants on the island remain sensitive to tem-perature change, and have different phenological characteristicsfrom natural forest and other cash forests such as lichee and longan(Chen et al., 2010). Rubber plantations have two distinct seasons ina year: a growing season and a non-growing season. In the growingseason, rubber plantations have high NDVI values, similar to otherevergreen forests. However, low temperatures in the winter(�10 �C) force rubber plants to defoliate, resulting in a canopy cov-erage of 20% (or less) and a low NDVI value (Chen et al., 2010).

We extracted the time series vegetation indices (NDVI, EVI andLSWI) from (1) one typical rubber site (Danzhou), which is thelong-term study site of the Rubber Research Institute, ChineseAcademy of Tropical Agricultural Sciences; (2) one evergreen forestsite in southwestern Hainan Island; and (3) one paddy rice site inthe Luodaixiang Town of Dongfang City (Fig. 3). The NDVI valuesof rubber plantation decreases substantially in the senescence per-iod due to cold and rubber tapping, which starts from late Octoberand ends in March of next year; and the lowest NDVI values appearin January, February and March. In the growing season, NDVI in-creases rapidly from the end of March to October (Fig. 3a). Thetemporal profiles of EVI and LSWI for rubber, paddy rice croplandand evergreen forests are different as well, but not as significantas NDVI (Fig. 3b and c). Therefore, we applied 250 m MOD13Q1NDVI products to identify and delineate phenology of rubber plan-tation. In this study we only identify and map mature rubber plan-tation, as young rubber is easily confused with bare soil or crops,especially when rubber seedlings are intercropped with othercrops.

The workflow for mapping rubber plantation is depicted inFig. 1. Rubber plantation was identified with the following decisionrule:

if ð0:5 < NDVIðJan;Feb;MarÞ < 0:7 \ 0:73 < NDVIðMay;Jun;JulÞ < 0:85Þ; rubber ¼ 1;

else; rubber ¼ 0

where NDVI(Jan, Feb, Mar) and NDVI(May, Jun, Jul) are the mean NDVI inJanuary, February and March and the mean NDVI in May, Juneand July, respectively. The monthly NDVIs were calculated withthe Maximum Value Composition (MVC) method from the 16-daycomposites in a month (most months have two 16-day compositedata, but there is only one 16-day composite in November). Theselection of the thresholds was basically based on a statisticalanalysis of 287 rubber plantation samples. For example, theNDVI(May, Jun, Jul) of rubber samples had a mean value of 0.79 and astandard deviation of 0.06, we set the thresholds as the mean valueminus and plus the standard deviation value; we further checkedthe data distribution and tested the thresholds, and finally verified0.73 and 0.85 as the thresholds of NDVI(May, Jun, Jul).

2.6. Accuracy assessment and comparison to other land cover datasets

We used both ground survey and Google Earth to evaluate theaccuracy of the land cover maps produced in this study. Firstly,we linked the PALSAR-based land cover map with Google Earthand randomly selected 40 polygons (31,284 pixels) as ground ref-erence from the Google Earth images for accuracy assessment. Inorder to assess the accuracy of the rubber plantation identification,we conducted a field survey in August of 2011 with a GPS digitalcamera as the main information capture tool. 87 georeferencedfield photos were acquired in the plots where rubber plantation

Fig. 3. Comparison between the temporal profiles of (a) NDVI, (b) EVI, and (c) LSWI for typical vegetation types (January 2006 to December 2008). The samples include (a) arubber sample in Danzhou station (19.516 N, 109.47 E), which is a typical observational site of Rubber Research Institute, Chinese Academy of Tropical Agricultural Sciences;(b) an evergreen forest sample (18.718 N, 108.966 E) in southwestern Hainan Island; and (c) a paddy rice cropland sample (19.058 N, 108.676 E) in the Luodaixiang Town ofDongfang City, Hainan. There are some gaps in the curves as the values with low quality (e.g. intensive cloud coverage or shadow) were removed.

24 J. Dong et al. / ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33

areas were larger than 500 m � 500 m. This data was also up-loaded into an online system named the ‘‘Global Geo-ReferencedField Photo Library’’ (http://www.eomf.ou.edu/photos/). Subse-quently, these plantation areas were outlined in Google Earthand used as ground truth polygons for evaluating the accuracy ofthe identification of rubber plantations.

We also compared the PALSAR-based land cover map with theLandsat-based National Land Cover Dataset (NLCD). The ChineseAcademy of Sciences has supported the development of NLCD in

China and developed a methodology based on visual interpreta-tion and digitalization of Landsat imagery (Liu and Deng, 2010;Liu et al., 2010). The NLCD vector dataset has a scale of1:100,000 (Liu et al., 2005). Four periods (the late 1980s, themid-1990s, 2000, and 2005) of NLCD datasets have been com-pleted, and they are comprehensive spatial datasets about landcover change at the national scale. In this study, we used theNLCD 2005 for comparison with the PALSAR-based land covermap.

J. Dong et al. / ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33 25

3. Results

3.1. Landcover map of forest, cropland, urban and water from PALSARin 2007

The signatures of four land cover types were explored in the HH,HV, Ratio image and Difference image based on the ROIs from Goo-gle Earth. The separability of these four land cover types was high(Fig. 4). Water has the lowest HH and HV values but higher Ratiovalues than the other three land cover types and is thus easilyidentified. Cropland has lower HV and HH values but higher Differ-ence values than that of forest. Higher HV values were observed forthe forests due to their large crown canopy which depolarizes inci-dent radiation. Urban land has higher HH and Difference valuesthan forest but lower Ratio values. However, urban land overlapswith forest to a large degree in all the four indicators due to com-plex reflectance environments in urban area, building orientations,corner reflectance, and large forest patches in urban areas.

The map produced with NN method is shown in Fig. 5a. Theareas of the four land cover types are 2.07 � 104 km2 forest;1.06 � 104 km2 cropland; 0.17 � 104 km2 urban land; and0.08 � 104 km2 water body, respectively. The forests are mainlyconcentrated in higher elevations of the central hilly area. Thecropland is distributed in a mosaic pattern in the lower elevations

1.8

1 6

1.8

1 4

1.6

1 2

1.4

% 1 0

1.2

ency

0 8

1.0

requ

e

0 6

0.8

Fr

0 4

0.6

0 2

0.4

0.2

0.0

-30 -20 -10 0 10

HH

1 81.8

1.6

1.4

%

1.2

ncy

% 1.0

quen 0.8

Fre

q

0.6

0.4

0.2

0.0

1 2 1 0 0 8 0 6 0 4 0 2 0 0 0 2 0 4 0 6 0 8 1 0

Ratio-1.2 -1.0 -0.8 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 1.0

(a)

(c)

Fig. 4. The histogram of the four land cover types in (a) PALSAR HH polarization, (b)difference image for all the pixels within the Region of Interest (ROIs) in Hainan Island,

surrounded by forests. This indicates that topography (Fig. 2) playsa role in the spatial pattern of forests and cropland on the island.The two large cities, Haikou and Sanya, account for the majorityof the urban land area. The PALSAR-based false color compositemap clearly illustrates land cover separability (Fig. 5c).

The overall accuracy of the land cover map was 89% with KappaCoefficient of 0.79 (Table 1). Both the Producer’s Accuracy and theUser’s Accuracy of forest and water were more than 92% (Table 1).Cropland also had a high Producer’s Accuracy (86%); however, theUser’s Accuracy of cropland was only 62%. The low User’s Accuracyshows that cropland area was underestimated in some places, asdifferent kinds of crops (e.g. fallowed cropland or different stagesof crops) have various polarization signatures. Urban land had aProducer’s Accuracy of 56% while its User’s Accuracy was 74%, assome trees in cities affected the pixels of urban land. In general,the classification results had reasonably good accuracy for the fourland cover types, especially forests.

A comparison between the PALSAR-based land cover map in2007 (PALSAR 2007) and the NLCD 2005 map was conductedand a confusion matrix was generated to evaluate the consistencyof the two land-cover classification results. Fig. 5a shows the spa-tial distribution of the PALSAR-based map, which is consistentwith the NLCD 2005 dataset (Fig. 5b). Forests have a relativelyhigh agreement; PALSAR-based forest area is 2.07 � 104 km2

2.52.5

2.0

%

1.5

ency

1 0

requ

e 1.0

Fr

0.5

0.0

-35 -30 -25 -20 -15 -10 -5 0HV

2 52.5

2.0 forest2.0

cropland

%

1.5 buildings

ncy

% water

quen 1.0

Fre

q

0 50.5

0.0

0 5 10 15 20 25 30

Difference0 5 10 15 20 25 30

(b)

(d)

PALSAR HV polarization, (c) PALSAR HH/HV ratio image, and (d) PALSAR HH–HVChina.

Fig. 5. A comparison between (a) PALSAR-based land cover map in 2007, (b) NLCD land cover map in 2005, and (c) PALSAR-based color composite map in 2007 (R/G/B = polarizations HH/HV/HH–HV Difference) in Hainan Island, China.

26 J. Dong et al. / ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33

while the forest in NLCD is 2.17 � 104 km2. The overlapped forestarea between these two datasets is 1.63 � 104 km2 (Table 2),which accounts for �79% of forest in the PALSAR 2007 map and�75% of forest in the NLCD 2005 dataset. The discrepancy be-tween these two datasets could be attributed partly to the landuse change from 2005 to 2007, and partly because the classifica-

tion systems differ slightly. The estimates of cropland area fromthese two land-cover classification schemes differ somewhat(Table 2). It is possible that cropland in the NLCD 2005 dataset in-cluded different cropland types, such as fallowed cropland, andthe classified cropland from PALSAR might have included sometall-grass grassland.

Fig. 6. (a) 50 m PALSAR-based per-pixel binary forest map; (b) 500 m forest area percentage map aggregated from the 50 m PALSAR-based forest map; (c) 500 m per-pixelbinary map of evergreen forest based on MODIS LSWI analysis; (d) area percentage map of evergreen forest by overlaying per-pixel binary evergreen forest map c and 500 mforest area percentage map b; (e) area percentage map of deciduous forest by subtracting evergreen forest area percentage map d from 500 m forest area percentage map b.

J. Dong et al. / ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33 27

3.2. The area and spatial distribution of evergreen and deciduousforests

The spatial distribution of forest area percentage within 500 mpixels is shown in Fig. 6b. The loss of forest area during the aggre-gation from 50 m spatial resolution (Fig. 6a) to 500 m and 250 m

spatial resolutions was negligible. The 500 m pixels including for-ests covered 94% of the whole island area.

The spatial distribution of evergreen forest area percentage isshown in Fig. 6d. The area of evergreen forest was estimated tobe approximately 0.75 � 104 km2. Those evergreen forest pixelswith high forest area percentage were concentrated in the middle

Table 2A comparison between PALSAR land cover map in 2007 and NLCD 2005 dataset.

NLCD 2005 (km2) PALSAR total (km2)

Forest Cropland Water Urban land Others

PALSAR 2007 Forest 16,268 3215 220 257 744 20,704Cropland 4382 5025 397 346 467 10,616Water 138 177 341 42 117 815Urban land 924 475 27 166 61 1654

NLCD total (km2) 21,712 8892 984 811 1389 33,789

Table 1A comparison between PALSAR land cover map and ground truth samples.

Ground truth samples (unit: pixels) PALSAR Total (pixel) User acc. (%)

Water Forest Cropland Urban land

PALSAR land cover classification Water 5074 6 193 0 5273 96.23Forest 0 18,839 7 1564 20,410 92.30Cropland 113 193 1245 468 2019 61.66Urban land 0 943 3 2636 3582 73.59

Total ground truth samples (pixel) 5187 19,981 1448 4668 31,284Prod. acc. (%) 97.82 94.28 85.98 56.47

28 J. Dong et al. / ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33

and southern regions with hilly topography, whereas the decidu-ous forests were mainly distributed in the northern parts of the is-land (Fig. 6e).

Fig. 7 shows the frequency distributions for forest percentagewithin 500 m and 250 m pixels for the evergreen forest pixels,deciduous forest pixels and rubber plantation pixels, and the re-gion as a whole. Most of the evergreen forest pixels have high for-est percentages; 49% of evergreen forest pixels have forestpercentages ranging from 95% to 100%. In other words, 49% ofMODIS-based evergreen forest pixels were almost pure evergreenforest. In addition, 70% of MODIS evergreen forest pixels have for-est area percentages greater than 65%, which revealed that ever-green forest distribution was concentrated to a great extent.However, deciduous forest showed different characteristics fromevergreen forest in that its distribution was more dispersed thanthat of evergreen forest. For example, 47% of 500 m deciduous for-est pixels have forest area percentages larger than 60%, 53% of500 m pixels have forest area percentages larger than 50%, andthe 500 m pixels with forest area proportion less than 5% accountsfor 10%.

1 According to ‘Outline of comprehensive land use planning of Hainan province006–2020)’ by Yansui Liu et al. (2008).

3.3. The area and spatial distribution of rubber plantation in 2007

The spatial distribution of rubber plantations from analysis ofMOD13Q1 (250 m) data in 2007 is shown in Fig. 8b. According tothe ground sampling based accuracy assessment the mappingaccuracy was high. The overall accuracy for rubber plantationswas 85% with kappa of 0.83.

The NDVI-based rubber plantation map has 1.04 � 105 pixels ofrubber plantations; it was overlaid with the PALSAR-based 250 mforest area percentage map in 2007 (Fig. 8a); and the resultantmap is assumed to be the map of rubber plantation in 2007(Fig. 8c). Approximately 43% of rubber plantation pixels (fromMOD13Q1, 250 m resolution) have a rubber plantation area per-centage of more than 95%, and 85% of rubber plantation pixels haverubber plantation area percentages larger than 50% (Fig. 7d). Thetotal area of rubber plantations was calculated by summing forestarea percentage over those rubber plantation pixels from theMOD13Q1 map (Fig. 8c), and it is estimated that there was a totalof 5149 km2 of rubber plantations on Hainan Island in 2007, which

is about 7% larger than the area estimate (4800 km2 in 2010) fromthe official statistical report (The Yearbook of Hainan Province in2011).

4. Discussion

This study evaluated the application potential of PALSAR 50 mmosaic and MODIS data for delineation and mapping of forests inthe tropical zone. The results were evaluated with ground truthdata and compared to the Landsat-based NLCD 2005 map gener-ated through visual interpretation. Our study revealed the advan-tages of PALSAR mosaic product as compared to Landsat images.The PALSAR-based forest area estimate (2.07 � 104 km2) is closerto the land survey result (2.02 � 104 km2) from the China Land Re-source Bureau1 than the Landsat-based NLCD-2005 data(2.17 � 104 km2). In addition to the potential land cover changesthat have taken place 2005–2007, these differences may be largelydue to methodological differences and image data. Note that theNLCD-2005 land cover map was generated by visual interpretationbased on color composition maps; forest and some crops could havesimilar optical (color) characteristics and are difficult to separatethem in human-aid interpretation on a false color composite map.With comparison in some selected regions by zooming in the map,we found that the PALSAR 2007 dataset has a higher accuracy thanthe NLCD-2005 dataset. The maximum ability for human vision todistinguish land cover from Landsat images is about 3 pixels (about100 m). Therefore, when compared to automated digital image clas-sification, the visual interpretation method omits some informationsometimes. The PALSAR 2007 forest map shows more spatial heter-ogeneity and has more specific details in the land parcels.

Before PALSAR, single HH polarization data from JERS-1 waswidely used for forest mapping (e.g. clear-cut or biomass intensity)(Luckman et al., 1998; Simard et al., 2002). However, this singlepolarization-based algorithm has some shortcomings in forestdelineation (Santoro et al., 2010). Multiple frequency, multiplepolarization and time series of SAR data improves classificationaccuracy evidently (Ranson and Sun, 1994; Townsend, 2002). Sev-

(2

Fig. 7. Frequency histograms of forest area proportion in (a) the entire Hainan Island, (b) evergreen forest pixels, (c) deciduous forest pixels, and (d) rubber plantation pixels.

J. Dong et al. / ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33 29

eral studies have evaluated the potential of PALSAR 50 m mosaicdata in forest identification and forest classification (Miettinen

and Liew, 2011; Thiel et al., 2009). In this study, we used not onlyHH and HV polarizations of PALSAR data, but also two new gener-

Fig. 8. (a) 250 m forest area percentage map aggregated from the 50 m PALSAR-based forest map; (b) 250 m per-pixel binary map of rubber plantation based on MODIS NDVIdata; (c) area percentage map of rubber plantation by combining the forest area percentage map a and the 250 m rubber plantation per-pixel binary map b.

30 J. Dong et al. / ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33

ated images (Ratio and Difference of HH and HV) that were provedeffective. However, more complicated forest type classificationmethods are less robust than a simple forest/non-forest classifica-tion (Walker et al., 2010).

Phenology of forests (deciduous and evergreen forests) in thetropical zone is affected by the seasonality of precipitation (dryseason and wet season). Hainan Island has a typical tropical mon-soon climate. Deciduous trees lose their leaves to conserve waterand avoid transpiration in the dry season (winter) and are so-calledtropical seasonal forest in this region. Evergreen trees grow all yearbut suffer greater water loss during the dry season, and they havedifferent photosynthetic rates and nutrient-use efficiency thandeciduous trees (Delucia and Schlesinger, 1995).

The pattern of forest composition (evergreen and deciduous for-ests) plays an important role in ecological processes and forest car-bon cycling. However, due to lack of time series of PALSAR 50 mmosaic data throughout a year, we were not able to map deciduousand evergreen forests from the PALSAR 50 m data. Instead, we usedtime series MODIS data to generate an evergreen forest map (per-

pixel map) and combined that map with the PALSAR-based forestfractional map to estimate evergreen and deciduous forests on Hai-nan Island. Note that time series imagery from optical sensors (e.g.,SPOT-VGT, AVHRR, MODIS) are widely used to generate regionaland global maps of evergreen and deciduous forests (DeFrieset al., 1995; Liang, 2001; Roderick et al., 1999). The leaf loss ofdeciduous trees is an important biophysical event to separate ever-green and deciduous forests. Time series data of LSWI were used toseparate evergreen and deciduous plants (Xiao et al., 2002), and tomap evergreen forests in the pan-tropical zone (Xiao et al., 2009). Acomparison between the PALSAR-based forest map and MODIS-based forest map will help evaluate the accuracy of the forestmap derived from MODIS imagery. The comparison between thePALSAR-based forest map and the MODIS-based evergreen forestmap in the Hainan Island showed that the LSWI-based approachfor mapping evergreen forest is robust.

Plantations, or cash forests, are an important land use type andeconomic activity. Miettinen and Liew (2011) found that the PAL-SAR 50 m mosaic product can be used to separate rubber, wattles

J. Dong et al. / ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33 31

and palms (oil palm and coconut combined) in known closed can-opy plantation areas. PALSAR 50 m mosaic data were used to iden-tify palm plantations in insular Southeast Asia with a method usingthreshold values (Miettinen et al., 2012). For Hainan Island, two re-cent studies have used a remote sensing approach to estimate thearea of rubber plantations (Zhang et al., 2010; Chen et al., 2010).One study used four Landsat TM images (acquired in May and June2008) and a supervised classification method to map rubber plan-tations on the island (Zhang et al., 2010); they estimated a totalarea of rubber plantations to be 4167.6 km2, which is about 8%lower than the estimate from the annual official statistical reportin 2008. The other study using maximum monthly NDVI compos-ites derived from the MODIS 16-day vegetation indices product(MOD13Q1) estimated a total area of rubber plantation to be4664 km2 on Hainan Island in 2009 (Chen et al., 2010).

Together with all the above mentioned studies, our study showsthe capabilities of remote sensing methods in forest monitoringand mapping of economically valuable cash plantations. It alsohighlights the benefits that can be achieved by combining severaldifferent types of remotely sensed data. Based on our findings dur-ing this study, we believe that the mapping of rubber plantationson Hainan Island could be further improved by combining or fusingthe time series of Landsat images and PALSAR 50 m mosaic imag-ery (Hong et al., 2009).

Finally, we would like to highlight the usability of online photoarchives for ground truth purposes. Evaluation (or validation) ofland cover maps is a major challenge in the community of remotesensing and land use and land cover change, as it often requires alarge number of ground truth data of individual land cover types.In this study, we used ground truth references data from bothgeo-referenced field photos acquired during field surveys andhigh-resolution images available in Google Earth. The GlobalGeo-Referenced Field Photos Library at the University of Oklahoma(http://www.eomf.ou.edu/photos/) allows users to share and ar-chive geo-referenced field photos; those field photos and associ-ated land cover database are downloaded in ‘kml’ or ESRI shapefiles online. As of January 2012, there are more than 35,000 highquality geo-referenced field photos in the database, which providesvaluable validation references for global land cover classification.In this study we also highlighted the Google Earth based validationworkflow. Google Earth’s High-Resolution Imagery Archive havehigh horizontal positional accuracy (Potere, 2008), especially in ur-ban areas, and is convenient for land cover validation in some re-gions. This method has been used in validation for land coverclassification (Benedek and Sziranyi, 2009; Cohen et al., 2010;Gemmell et al., 2009; Montesano et al., 2009). A specific tool thatintegrates Google Earth with image processing software synchro-nously could improve the convenience greatly in future.

5. Conclusion

A regional map of tropical forest distributions (e.g. evergreen ordeciduous) is imperative for ecological modeling and forest man-agement in pan-tropical regions (Achard et al., 2002). Mapping ofthe rubber plantation and its spatial distribution is also necessaryfor not only rubber industry but also regional development anddecision making (Li and Fox, 2011). Considering the cloud limita-tion from traditional optical remote sensing in pan-tropical re-gions, an integrating approach combining PALSAR 50 m andMODIS imagery was found to be alternative method for mappingforest and rubber plantation in the region, as this method incorpo-rates both cloud-free forest information from PALSAR and phenol-ogy information from MODIS. Evergreen and deciduous forestswere separated and rubber plantations identified using phenologyinformation based on MODIS temporal profile analysis. We foundthat the PALSAR 50 m Orthorectified Mosaic Product can be used

effectively to delineate forest, cropland, water body and urbanland. The sensitivity of rubber plants in Hainan Island to climatemakes the rubber separable from other types of forests. This studydemonstrated that an integrated strategy combining PALSAR andMODIS data could be more reliable in tropical forest classificationand cash forest (e.g. rubber plantation) identification.

Acknowledgements

This study was supported by the NASA Land Use and Land CoverChange Program (NNX09AC39G), the US National ScienceFoundation (NSF) EPSCoR Program (NSF-0919466), and theChinese National Key Program for Developing Basic Science(2010CB950900). We thank journal editor Dr. Daniel L. Civco andthree reviewers for their valuable suggestions and comments onearlier version of the manuscript.

Appendix A. Supplementary material

Supplementary data associated with this article can be found, inthe online version, at http://dx.doi.org/10.1016/j.isprsjprs.2012.07.004. These data include Google maps of the most impor-tant areas described in this article.

References

Achard, F., Estreguil, C., 1995. Forest classification of Southeast Asia using NOAAAVHRR data. Remote Sensing of Environment 54 (3), 198–208.

Achard, F., Eva, H., Mayaux, P., 2001. Tropical forest mapping from coarse spatialresolution satellite data: production and accuracy assessment issues.International Journal of Remote Sensing 22 (14), 2741–2762.

Achard, F., Eva, H.D., Stibig, H.J., Mayaux, P., Gallego, J., Richards, T., Malingreau, J.P.,2002. Determination of deforestation rates of the world’s humid tropicalforests. Science 297 (5583), 999–1002.

Achard, F., DeFries, R., Eva, H., Hansen, M., Mayaux, P., Stibig, H.J., 2007. Pan-tropicalmonitoring of deforestation. Environmental Research Letters 2 (4), 045022.

Ardila, J.P., Tolpekin, V., Bijker, W., 2010. Angular backscatter variation in L-bandALOS ScanSAR images of tropical forest areas. IEEE Transactions on Geoscienceand Remote Sensing 7 (4), 821–825.

Asner, G.P., 2001. Cloud cover in Landsat observations of the Brazilian Amazon.International Journal of Remote Sensing 22 (18), 3855–3862.

Asner, G.P., Knapp, D.E., Broadbent, E.N., Oliveira, P.J.C., Keller, M., Silva, J.N., 2005.Selective logging in the Brazilian Amazon. Science 310 (5747), 480–482.

Baghdadi, N., Boyer, N., Todoroff, P., El Hajj, M., Begue, A., 2009. Potential of SARsensors TerraSAR-X, ASAR/ENVISAT and PALSAR/ALOS for monitoring sugarcanecrops on Reunion Island. Remote Sensing of Environment 113 (8), 1724–1738.

Bartholome, E., Belward, A.S., 2005. GLC2000: a new approach to global land covermapping from Earth observation data. International Journal of Remote Sensing26 (9), 1959–1977.

Benedek, C., Sziranyi, T., 2009. Change detection in optical aerial images by amultilayer conditional mixed Markov model. IEEE Transactions on Geoscienceand Remote Sensing 47 (10), 3416–3430.

Bond-Lamberty, B., Peckham, S.D., Ahl, D.E., Gower, S.T., 2007. Fire as the dominantdriver of central Canadian boreal forest carbon balance. Nature 450 (7166), 89–92.

Cai, Y., 1994. Factor analysis and development stratege of land resource use inHainan Island. Natural Resources (2), 1–7.

Chen, H., Chen, X., Chen, Z., Zhu, N., Tao, Z., 2010. A primary study on rubber acreageestimation from MODIS-based information in Hainan. Chinese Journal ofTropical Crops 31 (07), 1181–1185.

Christian, H., Narong, C., 2005. Effects of Land Use Changes on Soil ChemicalProperties of Sandy Soils from Tropical Hainan, China, Management of TropicalSandy Soils for Sustainable Agriculture, Khon Kaen, Thailand.

Cohen, W.B., Yang, Z.G., Kennedy, R., 2010. Detecting trends in forest disturbanceand recovery using yearly Landsat time series: 2. TimeSync – Tools forcalibration and validation. Remote Sensing of Environment 114 (12), 2911–2924.

Collins, M.J., Dymond, C., Johnson, E.A., 2004. Mapping subalpine forest types usingnetworks of nearest neighbour classifiers. International Journal of RemoteSensing 25 (9), 1701–1721.

DeFries, R., Hansen, M., Townshend, J., 1995. Global discrimination of land covertypes from metrics derived from AVHRR pathfinder data. Remote Sensing ofEnvironment 54 (3), 209–222.

Delucia, E.H., Schlesinger, W.H., 1995. Photosynthetic rates and nutrient-useefficiency among evergreen and deciduous shrubs in Okefenokee Swamp.International Journal of Plant Sciences 156 (1), 19–28.

32 J. Dong et al. / ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33

Dixon, R.K., Brown, S., Houghton, R.A., Solomon, A.M., Trexler, M.C., Wisniewski, J.,1994. Carbon pools and flux of global forest ecosystems. Science 263 (5144),185–190.

Foley, J.A., DeFries, R., Asner, G.P., Barford, C., Bonan, G., Carpenter, S.R., Chapin, F.S.,Coe, M.T., Daily, G.C., Gibbs, H.K., Helkowski, J.H., Holloway, T., Howard, E.A.,Kucharik, C.J., Monfreda, C., Patz, J.A., Prentice, I.C., Ramankutty, N., Snyder, P.K.,2005. Global consequences of land use. Science 309 (5734), 570–574.

Fox, J., Vogler, J.B., 2005. Land-use and land-cover change in montane mainlandsoutheast Asia. Environmental Management 36 (3), 394–403.

Friedl, M.A., McIver, D.K., Hodges, J.C.F., Zhang, X.Y., Muchoney, D., Strahler, A.H.,Woodcock, C.E., Gopal, S., Schneider, A., Cooper, A., Baccini, A., Gao, F., Schaaf, C.,2002. Global land cover mapping from MODIS: algorithms and early results.Remote Sensing of Environment 83 (1–2), 287–302.

Gemmell, A.L., Smith, G.C., Haines, K., Blower, J.D., 2009. Validation of ocean modelsyntheses against hydrography using a new web application. Journal ofOperational Oceanography 2 (2), 29–41.

Giri, C., Zhu, Z.L., Reed, B., 2005. A comparative analysis of the Global Land Cover2000 and MODIS land cover data sets. Remote Sensing of Environment 94 (1),123–132.

Hong, G., Zhang, Y., Mercer, B., 2009. A wavelet and IHS integration method to fusehigh resolution SAR with moderate resolution multispectral images.Photogrammetric Engineering & Remote Sensing 75 (10), 1213–1223.

Huang, C.Q., Goward, S.N., Schleeweis, K., Thomas, N., Masek, J.G., Zhu, Z.L., 2009.Dynamics of national forests assessed using the Landsat record: case studies ineastern United States. Remote Sensing of Environment 113 (7), 1430–1442.

Huang, C., Goward, S.N., Masek, J.G., Thomas, N., Zhu, Z., Vogelmann, J.E., 2010. Anautomated approach for reconstructing recent forest disturbance history usingdense Landsat time series stacks. Remote Sensing of Environment 114 (1), 183–198.

Huete, A.R., Liu, H.Q., Batchily, K., vanLeeuwen, W., 1997. A comparison ofvegetation indices over a global set of TM images for EOS-MODIS. RemoteSensing of Environment 59 (3), 440–451.

Huete, A., Didan, K., Miura, T., Rodriguez, E.P., Gao, X., Ferreira, L.G., 2002. Overviewof the radiometric and biophysical performance of the MODIS vegetationindices. Remote Sensing of Environment 83 (1–2), 195–213.

Hunan, S.O., 1999. The Past 50 years of Hainan (1949–1999). China Statistics Press.Kabir, S., He, D.C., Sanusi, M.A., Hussin, W.M.A.W., 2010. Texture analysis of IKONOS

satellite imagery for urban land use and land cover classification. ImagingScience Journal 58 (3), 163–170.

Kavzoglu, T., 2009. Increasing the accuracy of neural network classification usingrefined training data. Environmental Modelling & Software 24 (7), 850–858.

Kummer, D.M., Turner, B.L., 1994. The human causes of deforestation in Southeast-Asia. Bioscience 44 (5), 323–328.

Lelieveld, J., Butler, T.M., Crowley, J.N., Dillon, T.J., Fischer, H., Ganzeveld, L., Harder,H., Lawrence, M.G., Martinez, M., Taraborrelli, D., Williams, J., 2008.Atmospheric oxidation capacity sustained by a tropical forest. Nature 452(7188), 737–740.

Li, Z., Fox, J.M., 2011. Integrating Mahalanobis typicalities with a neural network forrubber distribution mapping. Remote Sensing Letters 2 (2), 157–166.

Li, Z., Fox, J.M., 2012. Mapping rubber tree growth in mainland Southeast Asia usingtime-series MODIS 250 m NDVI and statistical data. Applied Geography 32 (2),420–432.

Li, H.M., Aide, T.M., Ma, Y.X., Liu, W.J., Cao, M., 2007. Demand for rubber is causingthe loss of high diversity rain forest in SW China. Biodiversity and Conservation16 (6), 1731–1745.

Liang, S., 2001. Land-cover classification methods for multi-year AVHRR data.International Journal of Remote Sensing 22 (8), 1479–1493.

Lin, M., Zhang, Y., 2001. Dynamic change of tropical forest in Hainan Island.Geographical Research 20 (06), 703–712.

Liu, J.Y., Deng, X.Z., 2010. Progress of the research methodologies on the temporaland spatial process of LUCC. Chinese Science Bulletin 55 (14), 1354–1362.

Liu, J.Y., Liu, M.L., Tian, H.Q., Zhuang, D.F., Zhang, Z.X., Zhang, W., Tang, X.M., Deng,X.Z., 2005. Spatial and temporal patterns of China’s cropland during 1990–2000: an analysis based on Landsat TM data. Remote Sensing of Environment 98(4), 442–456.

Liu, J.Y., Zhang, Z.X., Xu, X.L., Kuang, W.H., Zhou, W.C., Zhang, S.W., Li, R.D., Yan, C.Z.,Yu, D.S., Wu, S.X., Nan, J., 2010. Spatial patterns and driving forces of land usechange in China during the early 21st century. Journal of Geographical Sciences20 (4), 483–494.

Longepe, N., Rakwatin, P., Isoguchi, O., Shimada, M., Uryu, Y., Yulianto, K., 2011.Assessment of ALOS PALSAR 50 m orthorectified FBD data for regional landcover classification by support vector machines. IEEE Transactions onGeoscience and Remote Sensing 49 (6), 2135–2150.

Luckman, A., Baker, J., Honzak, M., Lucas, R., 1998. Tropical forest biomass densityestimation using JERS-1 SAR: seasonal variation, confidence limits, andapplication to image mosaics. Remote Sensing of Environment 63 (2), 126–139.

Miettinen, J., Liew, S.C., 2011. Separability of insular Southeast Asian woodyplantation species in the 50 m resolution ALOS PALSAR mosaic product. RemoteSensing Letters 2 (4), 299–307.

Miettinen, J., Shi, C., Tan, W.J., Chinliew, S., 2012. 2010 land cover map of insularSoutheast Asia in 250-m spatial resolution. Remote Sensing Letters 3 (1), 11–20.

Montesano, P.M., Nelson, R., Sun, G., Margolis, H., Kerber, A., Ranson, K.J., 2009.MODIS tree cover validation for the circumpolar taiga–tundra transition zone.Remote Sensing of Environment 113 (10), 2130–2141.

Morton, D.C., DeFries, R.S., Shimabukuro, Y.E., Anderson, L.O., Espirito-Santo, F.D.B.,Hansen, M., Carroll, M., 2005. Rapid assessment of annual deforestation in theBrazilian Amazon using MODIS data. Earth Interactions 9 (8), 1–22.

Page, S.E., Siegert, F., Rieley, J.O., Boehm, H.D.V., Jaya, A., Limin, S., 2002. The amountof carbon released from peat and forest fires in Indonesia during 1997. Nature420 (6911), 61–65.

Perea, A.J., Merono, J.E., Aguilera, M.J., de la Cruz, J.L., 2010. Land-cover classificationwith an expert classification algorithm using digital aerial photographs. SouthAfrican Journal of Science 106 (5–6), 82–87.

Pielke, R.A., 2005. Land use and climate change. Science 310 (5754), 1625–1626.Potere, D., 2008. Horizontal positional accuracy of Google earth’s high-resolution

imagery archive. Sensors 8 (12), 7973–7981.Qiu, J., 2009. Where the rubber meets the garden. Nature 457 (7227), 246–247.Ranson, K.J., Sun, G.Q., 1994. Northern forest classification using temporal

multifrequency and multipolarimetric Sar images. Remote Sensing ofEnvironment 47 (2), 142–153.

Richards, J.A., Jia, X., 2006. Remote Sensing Digital Image Analysis: An Introduction,fourth ed. Springer, Berlin.

Roderick, M.L., Noble, I.R., Cridland, S.W., 1999. Estimating woody and herbaceousvegetation cover from time series satellite observations. Global Ecology andBiogeography 8 (6), 501–508.

Rosenqvist, A., Shimada, M., Ito, N., Watanabe, M., 2007. ALOS PALSAR: a pathfindermission for global-scale monitoring of the environment. IEEE Transactions onGeoscience and Remote Sensing 45 (11), 3307–3316.

Sakaguchi, K., Zeng, X.B., Christoffersen, B.J., Restrepo-Coupe, N., Saleska, S.R.,Brando, P.M., 2011. Natural and drought scenarios in an east central Amazonforest: fidelity of the Community Land Model 3.5 with three biogeochemicalmodels. Journal of Geophysical Research 116, G01029.

Santoro, M., Fransson, J.E.S., Eriksson, L.E.B., Ulander, L.M.H., 2010. Clear-cutdetection in Swedish boreal forest using multi-temporal ALOS PALSARbackscatter data. IEEE Journal of Selected Topics in Applied EarthObservations and Remote Sensing 3 (4), 618–631.

Simard, M., Saatchi, S.S., De Grandi, G., 2000. The use of decision tree and multiscaletexture for classification of JERS-1 SAR data over tropical forest. IEEETransactions on Geoscience and Remote Sensing 38 (5), 2310–2321.

Simard, M., De Grandi, G., Saatchi, S., Mayaux, P., 2002. Mapping tropical coastalvegetation using JERS-1 and ERS-1 radar data with a decision tree classifier.International Journal of Remote Sensing 23 (7), 1461–1474.

Stibig, H.J., Malingreau, J.P., 2003. Forest cover of insular southeast Asia mappedfrom recent satellite images of coarse spatial resolution. Ambio 32 (7), 469–475.

Stibig, H.J., Achard, F., Fritz, S., 2004. A new forest cover map of continentalsoutheast Asia derived from SPOT-VEGETATION satellite imagery. AppliedVegetation Science 7 (2), 153–162.

Su, W., Zhang, C., Yang, J.Y., Wu, H.G., Chen, M.J., Yue, A.Z., Zhang, Y.N., Sun, C., 2010.Knowledge-based object oriented land cover classification using SPOT5 imageryin forest-agriculture ecotones. Sensor Letters 8 (1), 22–31.

Thessler, S., Sesnie, S., Bendana, Z.S.R., Ruokolainen, K., Tomppo, E., Finegan, B., 2008.Using k-nn and discriminant analyses to classify rain forest types in a LandsatTM image over northern Costa Rica. Remote Sensing of Environment 112 (5),2485–2494.

Thiel, C.J., Thiel, C., Schmullius, C.C., 2009. Operational large-area forest monitoringin Siberia using ALOS PALSAR summer intensities and winter coherence. IEEETransactions on Geoscience and Remote Sensing 47 (12), 3993–4000.

Torbick, N., Salas, W.A., Hagen, S., Xiao, X.M., 2011. Monitoring rice agriculture inthe Sacramento Valley, USA with multitemporal PALSAR and MODIS imagery.IEEE Journal of Selected Topics in Applied Earth Observations and RemoteSensing 4 (2), 451–457.

Tottrup, C., Rasmussen, M.S., Eklundh, L., Jonsson, P., 2007. Mapping fractional forestcover across the highlands of mainland Southeast Asia using MODIS data andregression tree modelling. International Journal of Remote Sensing 28 (1–2),23–46.

Townsend, P.A., 2002. Estimating forest structure in wetlands using multitemporalSAR. Remote Sensing of Environment 79 (2–3), 288–304.

Townshend, J.R.G., Justice, C.O., 1988. Selecting the spatial-resolution of satellitesensors required for global monitoring of land transformations. InternationalJournal of Remote Sensing 9 (2), 187–236.

Tucker, C.J., 1979. Red and photographic infrared linear combinations formonitoring vegetation. Remote Sensing of Environment 8 (2), 127–150.

Walker, W.S., Stickler, C.M., Kellndorfer, J.M., Kirsch, K.M., Nepstad, D.C., 2010.Large-area classification and mapping of forest and land cover in the BrazilianAmazon: a comparative analysis of ALOS/PALSAR and Landsat data sources. IEEEJournal of Selected Topics in Applied Earth Observations and Remote Sensing 3(4), 594–604.

Wu, F., Wang, C., Zhang, H., Zhang, B., Tang, Y.X., 2011. Rice crop monitoring inSouth China with RADARSAT-2 quad-polarization SAR data. IEEE Transactionson Geoscience and Remote Sensing 8 (2), 196–200.

Xiao, X.M., Boles, S., Liu, J.Y., Zhuang, D.F., Liu, M.L., 2002. Characterization of foresttypes in Northeastern China, using multi-temporal SPOT-4 VEGETATION sensordata. Remote Sensing of Environment 82 (2–3), 335–348.

Xiao, X.M., Hollinger, D., Aber, J., Goltz, M., Davidson, E.A., Zhang, Q.Y., Moore, B.,2004. Satellite-based modeling of gross primary production in an evergreenneedleleaf forest. Remote Sensing of Environment 89 (4), 519–534.

Xiao, X.M., Zhang, Q.Y., Hollinger, D., Aber, J., Moore, B., 2005. Modeling grossprimary production of an evergreen needleleaf forest using modis and climatedata. Ecological Applications 15 (3), 954–969.

J. Dong et al. / ISPRS Journal of Photogrammetry and Remote Sensing 74 (2012) 20–33 33

Xiao, X., Biradar, C., Czarnecki, C., Alabi, T., Keller, M., 2009. A simple algorithm forlarge-scale mapping of evergreen forests in tropical America, Africa and Asia.Remote Sensing 1 (3), 355–374.

Xiao, W., Wang, X., Ling, F., 2010. The application of ALOS PALSAR data on mangroveforest extraction. Remote Sensing Technology and Application 25 (01), 91–96.

Xie, C., Li, Z., Li, X., 2010. A study of deformation in permafrost regions of Qinghai-Tibet Plateau based on ALOS/PALSAR D-InSAR interferometry. Remote Sensingfor Land & Resources 30 (01), 53–56.

Xu, X.L., Zeng, L., Zhuang, D.F., 2002. Analysis on land-use change and socio-economic driving factors in Hainan Island during 50 years from 1950 to 1999.Chinese Geographical Science 12 (3), 193–198.

Yang, Y., Li, Z., Chen, e., Ling, F., 2010. Recognition of forest cover based on multi-temporal dual polarization ALOS PALSAR data. Journal of Anhui AgriculturalSciences 38 (18), 9840–9844.

Zhang, J., Tao, Z., Liu, S., Cai, D., Tian, G., Xie, R., Xu, X., 2010. Rubber planting acreagecalculation in Hainan Island based on TM image. Chinese Journal of TropicalCrops 31 (04), 661–665.

Ziegler, A.D., Fox, J.M., Xu, J.C., 2009. The rubber juggernaut. Science 324 (5930),1024–1025.